OpenBind-0

Co-folding models predict a protein structure and place a small molecule inside it in one pass, and the benchmarks that rank them are drawn largely from the Protein Data Bank's back catalogue — complexes whose close relatives the models have usually already seen. Point the same models at a target a medicinal chemistry campaign is actively working on and the numbers change character: on the RNA-dependent RNA polymerase (RdRp) domain of NS5 from dengue serotype 2 and Zika virus, no co-folding method tested reached a 10% success rate.

OpenBind-0 (OB0) is the first model from the OpenBind Consortium, a UK-led open science effort based at Diamond Light Source generating protein-ligand crystal structures at scale, built with the AlQuraishi Lab at Columbia University and the OpenFold Consortium. It uses a nearly final version of the forthcoming OpenFold3 architecture, trained on PDB structures released through 30 June 2025, with checkpoints selected to optimise protein-ligand accuracy while retaining performance on other modalities. It also brings inference-time chemical steering to the OpenFold3 family, improving the physical plausibility of predicted ligand geometry rather than just its position.

OB0 is deliberately a baseline rather than a showcase: its training set contains no OpenBind data, only public structures, so that later models in the series — each meant to stay within roughly six months of the latest public structural data — can be measured against it as consortium-generated data enters the mixture. It was released on 21 August 2026 with its code and weights under Apache 2.0, alongside 717 ligand-bound crystal structures from three discovery campaigns.

#Key Features

• Protein-ligand-selected checkpoint: Checkpoint selection optimised for protein-ligand performance, unlike the general-purpose OpenFold3 preview weights it supersedes.
• Inference-time chemical steering: Steering during diffusion sampling, in the manner of Boltz and Protenix, raises the joint rate of correct pose and PoseBusters-valid geometry from 48% to 61%.
• Recent training cutoff: PDB data through 30 June 2025, versus September 2021 for OpenFold3-preview2 and AlphaFold3, with a stated cadence of periodic re-training.
• Open release: Apache 2.0 for the code and the weights and CC BY 4.0 for the training and evaluation data, with weights downloading without registration from the public openfold3-data S3 bucket.
• Generic training recipe, not the OB0 run: The training loop, project and DeepSpeed configs and the training guide are published under Apache 2.0, but the specific OpenBind-0 run configuration is not in the repository.
• Companion benchmark data: A 462-system Runs N' Poses-style benchmark with ground-truth structures and MSAs, plus 717 newly released fragment and hit binding events.

#Technical Details

OB0 pairs a representation trunk with a diffusion module that emits raw atom coordinates, following the AlphaFold3 design that OpenFold3 reproduces; the two share similar though not identical architectures. It is distributed as the default parameter set of the openfold3 Python package from version 0.5.0, as checkpoint openbind-2025-06-30-174k (of3-ob-2025-06-30-174k.pt), trained for 174,000 steps and fetched by setup_openfold. The same release moved the diffusion transformer's pair layer norm out of attention pair bias to match the AlphaFold3 supplement and fixed several MSA featurisation bugs. One caveat for anyone reproducing the numbers: the inference-time chemical steering behind the headline 61% joint success rate has no implementation in the public repository, so the steered arm cannot be rerun from the released code.

Evaluation follows the Runs N' Poses protocol on complexes released after the training cutoff, scoring a prediction correct at lDDT-PLI above 0.8 and ligand RMSD under 2 Å. Across similarity bins OB0 is comparable to AlphaFold3, Boltz-2 and Protenix. A notable negative result: four extra years of PDB data move most benchmark complexes into the same similarity bin as before, and OB0 only clearly beats OpenFold3-preview2 on the complexes whose bin actually changed. On the newly released campaign data, OB0 reaches 92.2% top-25 success on EV-A71 2A protease follow-ons (having seen the corresponding fragment structures in pre-training) and 28.2% on fatty acid thioesterase A, where Protenix-v1-20250630 scores higher at 39.4% and OpenFold3-preview2 reaches 14.5%. On the two RdRp systems every model tops out at 7.7% (DENV-2) and 6.7% (ZIKV).

#Applications

OB0 targets structure-based drug discovery: predicting how fragments and follow-on compounds bind, prioritising which analogues to make, and generating poses clean enough to feed free-energy calculations. Because it is permissively licensed and retrainable, discovery teams can fine-tune it on proprietary structures while retaining examples from the original training distribution — what the consortium notes makes fine-tuning robust. Within OpenBind it also closes an active-learning loop, with predictions helping select which crystallography experiments to run next.

#Impact

OB0's contribution is as much methodological as it is a model. By shipping discovery-relevant benchmarks — targets where success rates fall below 10%, with a failure analysis showing models miss the pocket entirely on flexible receptors — it gives the field evidence about where co-folding stands in drug development rather than on curated retrospective sets, and it establishes the control against which the consortium's future data releases will be judged. The caveats are real: the release is documented by a blog post and release notes rather than a peer-reviewed paper, performance varies widely by target, target-specific fine-tuning helped far less than pre-training on relevant structures, and the accompanying benchmark set is a preview of an unreleased PLINDER version that its authors ask users to replace once the official set lands.